It is one of the biggest problems in modern biology. The news that Deepmind’s A.I has shown extremely high accuracy results on the protein folding problem will accelerate Bioinformatics and biology research. The implications for drug discovery are enormous.
I know that people say this but I have questions:
1. Are there concrete examples where drugs have been discovered from protein folding structures? What are the biggest ones?
2. Is there machinery that already exists to take in 3D protein structures and create drugs? or is this yet another issue?
3. How does folding of a protein in the current state impacting the use of the protein when it is used? Presumably these proteins are similar to polymers where they are not super rigid in all environments, how does the environment effect the protein folding?
Two of the first, back in 1999 [1], were Zanamivir and Oseltamivir (Tamiflu). Influenza neuraminidase inhibitors. Researchers examined a structure of neuraminidase co-crystallized with its substrate and designed a sialic acid paralog that was designed to bind with residues that are more conserved across different known sequences of influenza.
I found a recent review with some others listed here [2]. It has a nice overview of the process too!
Forgot to answer your other questions. I'm not up to date on the structure-based drug design workflow but back when I did similar work (5 years ago) there definitely were rudimentary systems for generating molecules and docking them. It may have improved significantly since then. But I would probably characterize it as a problem it itself for sure.
Your other question is a VERY good one. Proteins usually fold into whatever may be most favorable based on the sequence, and it mostly stays consistent once it does. However, they are very flexible and structures solved by EM or x-ray crystallography are like a photograph of bird flapping its wings: you will see the wings in a position, and if you happen to have a few birds in the photograph, you might get a sense of where those wings can move to, but it's never going to be perfect. But like wings, proteins usually still have a limited amount of movement. There are other types that are much harder to understand that have less structure, but globular proteins that bind to drugs like this are usually pretty well-predicted by the snapshots we can get.
For docking: Autodock Vina (from Scripps) is the most frequently cited docking software in the biomedical research literature. It's open source.
Researchers use docking software to run libraries of existing drugs as well as design never-seen-before drugs out of the enzyme protein's active-site pocket.
These operations have been performed extensively this year (by research groups all over the world) on covid's main protease enzyme as well as the spike-ACE2 interface, for example.
[followup] and so frankly, it's hard to imagine a world where drug discovery isn't enormously sped up by an automated protein-folding approach which docking software like Autodock Vina require to be run. I know that not all of the pharma industry agrees with this assertion however...:
https://twitter.com/michael_gilman/status/133375535280704307...
My take: Since 0.1% of proteins whose amino acids have been sequenced have ever seen a crystal structure (i.e. the folded model) generated of them. an automated approach to 3D model generation
1) will have enormous implications on drug development, and 2) will most likely come from a new and very different generation of drug developers, who don't have a lot in common with the generation that produced the tweet pasted above.
My take is exactly the opposite: since 3D structures of proteins alone are almost never the bottleneck in drug discovery, this won't actually change anything. Knowing how the drug is going to bind, and knowing how it'll behave in vivo, are not something you can predict from deep sequencing data.
Some proteins require help to fold properly (because there are more than one enrgetically favored conformations). The helping enzymes are called chaperons.
A known 3D structure for your target protein is very useful to improve molecules that bind to it, but we can't yet determine which molecules bind to a target without actually trying it experimentally. Of course there are methods to predict binding, but they not reliable enough and in the end the drug candidates are discovered by throwing a lot of molecules at a specific target or assay.
Once you have a candidate, it is very useful to determine the structure of the protein together with the drug candidate. There you can see how it binds, and can make some educated guesses on how to change the molecule to make it bind better, or to improve other aspects without making it bind worse.
Determing the protein fold from scratch without experimental data is impressive, but it doesn't have an immediate use for drug development. But a few steps further and it could certainly help if you can also predict which molecules bind to the protein structure.
I would strongly recommend the following blog post from Derek Lowe to put the importance of this into context for drug development:
The other point missing in most of these discussions is we already know how most drug targets fold, even if we don't know the exact structure at atomic detail. It's everything else about their structure, dynamics, and in vivo function that remains very difficult. The real promise in AlphaFold IMHO isn't that we can magically solve protein structures without experiments (most really interesting structures are beyond what it can do anyway), but the more general application of these AI methods to human health.
For the most part, people don't even try to make drugs for proteins that don't have structures, so that's one. As was mentioned in the anecdote, even as it exists AlphaFold can be an extremely powerful ally in structure elucidation in combination with lab methods. So this will help us increase the targetable list of proteins, especially the tricky ones that were harder to crystallize.
Once Alpha Fold or future programs get better with side chain modeling (not even for the entire protein just some parts), they will also allow complete computer based design of new antibodies against any target of choice (this is currently only possible through experiments and the technologies that allow this are all heavily patented and proprietary).
Variations of AlphaFold will also be significantly useful in research in general, potentially becoming fundamental enough that every project working with proteins might reach out to this tool like they reach out to to say mass spectrometry or flow cytometry.
It's one stage, and it's an important stage, but it's not the most important stage for speeding up discovery. Most of the wasted time is due to developing drugs for what look to be promising targets, taking them into the clinic, and finding that they don't actually do anything to the disease even if they're perfect at affecting the target in the desired way.
Protein folding will help develop candidates faster. That's good. But it won't seriously help find the right targets faster, so I don't expect a substantial speedup in overall drug development times.
Dynamic modeling of proteins is an incredibly resource intensive process, and finding a ligand that binds to a model can also be difficult if the model is wrong.
What the folding model essentially does, is confirm that the modeled folded protein has the correctly modeled binding energies for each individual atom. Making modeling induced fit, and conformationally dynamic and difficult to drug proteins, more easy to model and find the correct ligands that bind to them.
Right. It makes creating a candidate drug to hit your target vastly easier. It doesn't really make picking the right target in the first place any easier. The preclinical stage will be faster. The clinical stage is still necessary, and that's the stage where the big bottlenecks (toxicity & efficacy) happen. There's not yet a reliable toxicity model, and efficacy is even harder.
This will certainly help. But we shouldn't expect drugs to be discovered in weeks instead of years, since the bit that usually takes a year or more (the clinical trials) isn't changed here.
None of that matters because all preclinical methods of testing for drug efficacy are not the same. For most of the important diseases of today, preclinical models are bstshit poor in predicting actual clinical efficacy. Mainly cancer.
> The implications for drug discovery are enormous.
As every disease has a mechanism via a protein channel, and all cells are made of proteins, doesn't this open the possibility of curing nearly any disease for which we understand it's mechanism and for which we can conceive of a protein shape to block or rebuild the tissue damages caused by such a disease? Delivery will still be an issue, but I don't see how this couldn't be used to prevent metastasis in tumors by custom created blocking proteins, for example.
For sure, it is probably just another perspective that has not being discovered yet that is why it seems like black magic. For example, we already know Muon And Neutrino particles are just flying around with very little interaction with our 3D world.
https://en.wikipedia.org/wiki/Antarctic_Muon_And_Neutrino_De...
DeepMind actually solved harder problems than most of the ones mentioned. All the proteins mentioned in the article (except for the chaperones) are 100 amino-acid and shorter, the therapeutic ones are even shorter, dozen(s) of AA. Thus, great preview of what is to come, though its from (2017). C&EN is truly the quantamagazine of biochem...
I wouldn't call quanta overhyped. Implications are at best embellished very modestly for entertainment purposes, but they do a very good job of staying true to the original paper and communicating some of the main ideas in a less-dumbed-down version than SA, and less ridiculously inaccurate/overhyped way than say physorg.
Maybe the physics/math editor/writers are better than the biology. I don't read the bio articles, so I can't comment directly about them. I just read the compsci/physics/math and they're not that bad.
To avoid a misconception, it is not known how a protein fold (as a verb) what is known, is how the final product is folded. The actual folding process may involve other molecules to assist the biolocical process
We do know quite a bit about how proteins fold in the cell, though of course there is likely still a lot to discover there. Chaperones are known for a long time now, and there's been plenty of research on them.
That's like saying "it's not known how stars form" because obviously we have not lived for a billion years to see a star form fully. We haven't seen a single Protein molecule fold fully, but we're pretty sure we know how it does (though some details are kinda murky).
my ephasis should be that much of "protein folding" is reffering to the final folded conformation and not the biological process of the protein actually folding
